GD Project a small lab of AI-native apps made with Godot

How it thinks

Not a folder of files.
A memory that understands.

Ask the way you think, in any language you like. Second Brain finds what you meant, even when you wrote it in other words. Try it on this small sample Brain of 29 notes.

Exact words–

Finds a note only if it contains the very words you typed, spelled the same way.

    By meaning · Second Brain–

    Compares what the question means with what each note means.

      A simulation in this page: nothing is sent. The sample Brain uses a small hand-made map of ideas. In the app, every note and every page of your documents gets a real meaning vector from an embeddings model, and a hit counts from a score of 0.25.

      One capture, end to end

      What happens between “add this” and the receipt

      Speak, drop a PDF, send a web page or a photo. The Archivist reads it, understands it, files it, indexes it twice and links it to what you already know. Pick something to capture and follow it through.

      1. Captureyou
      2. Readhear · extract · see
      3. Understandthe Archivist
      4. FilePARA
      5. Indexwords + meaning
      6. Linksame ideas
      7. ReceiptUndo
      –

      –

      –

      
                  

      The meaning map

      Notes that mean the same thing find each other

      Each dot is a note, placed by what it means, not by where it was filed. Notes about the same idea drift together, across places. The dashed lines are the “same idea” links the Brain suggests; you keep them or dismiss them.

      Under the hood

      Three layers of knowing, in one file on your computer

      Your Brain is a single SQLite database in your user folder. Every item lives there with three ways to be found, next to its original file and the history of every change.

      1

      Words

      A full-text index ranks notes and document pages by the words they hold (BM25). A word in the title counts four times as much as in the body. Exact names, numbers and codes are found here.

      notes_fts · chunks_fts
      2

      Meaning

      Every note, and every page of every document cut into pieces of about 1,200 characters, gets a vector of 1,536 numbers that captures what it means. Same idea in other words, even in another language: close vectors.

      note_embeddings · source_chunks

      Around them: places (PARA), types and fields (a recipe has servings), tags, originals of your files, restore points and a timeline of every change. spaces · item_types · sources · restore_points

      How a question is answered

      1. Meaning first. The question becomes a vector; the closest notes above 0.25 come back, best first.
      2. Words complete it. Keyword hits not already found fill the list: an exact name is never lost.
      3. Documents next. For “what does the quote say?”, the same two searches run on document pieces, and the answer cites the page.
      4. Places when you name one. “What's in my Inbox?” lists the place itself, newest first.

      Compared with a folder of text files and an AI that reads them

      A folder of text filesSecond Brain GD
      You wrote “quartz for the island”, you search “countertop”Not foundFound by meaning
      Related notesThe links you typed by handSuggested for every note, kept or dismissed by you
      A scanned PDFAn image nobody can searchRead page by page (local OCR), searchable, cited by page
      FilingYou pick the folderThe Archivist files in PARA and asks when it isn't sure
      An AI looking for an answerOpens and reads files until it finds itSearches with tools, reads only the few items that matter
      An AI that writesEdits your files directlyProposes; the Archivist files with your rules; a receipt and Undo
      Going backIf you set up versioningRestore points and a timeline of every change, built in

      AI-native

      The right model for each job

      Second Brain is not one chatbot. Each job goes to the model that does it best: a fast one to talk, a reasoning one for the weekly review, a transcriber for your voice, an embeddings model for meaning, and your own computer for what it can do alone. Pick a way in to see the route.

      You

      The engine

      Hearingspeech to text
      Readingtext, pages, pictures
      Archivistfiles, links, answers, tools
      Memorywords + meaning
      Voicespeaks back

      Models

      gpt-realtime-whisperdictation, live transcription
      gpt-realtime-2.1-minilive conversation
      gpt-6-solchat with tools, filing, agents, missions
      gpt-6-astrareasoning: weekly review, syntheses
      gpt-4o-minivision, hosted web search
      text-embedding-3-smallmeaning vectors
      gpt-4o-mini-ttsspoken replies
      This computerPDF text, OCR, search, graph

      Pick a way in on the left.

      Measured, not guessed

      Chat with tools moved to gpt-6-sol after replaying the owner's real failures on a throwaway Brain: 12 out of 12, against 11 for gpt-4o. Filing: 10 out of 10 test captures in 1.45 s.

      Never stuck

      Every model is a setting. When your OpenAI project lacks one, the job falls back (to gpt-4o or gpt-4o-mini) instead of failing.

      Local where it can be

      PDF text, scanned pages (Windows OCR or Tesseract), keyword search, the graph and your files stay on your computer. Local AI models are on the roadmap.

      Agents

      An AI that searches, instead of reading everything

      A Brain grows to thousands of items; no model should swallow it whole. The agents use tools, like you would use a search engine: search, open the few items that matter, cite them, act only through tools you can undo.

      Archivistready
      A

      Archivist

      Agent number one. Files, links, answers with sources, drives the app by voice. About sixty tools, every write undoable.

      S

      Sparring

      Your thinking partner. Reads your Brain to question, challenge or reassure you; changes nothing, and saves a synthesis only when you approve it.

      B

      Agent Builder

      Interviews you, writes a new agent's instructions and picks its tools. You see the draft; it exists when you say yes.

      Over MCP

      One memory for every agent you use

      Claude Code, Codex or Claude Desktop plug into the Brain over MCP. They search it by meaning; what they learn comes back as proposals the Archivist files with your rules. What one agent learns, the next one knows.

      A loop in five steps: an agent asks, the Brain answers by meaning with sources, the agent proposes what it learned, the Archivist files it with your rules, and the next session finds it.
      1. Ask. A Claude Code session in Agent Manager calls search: “what do we know about the quote?”
      2. Answer. The Brain searches by meaning and words, and sends three short hits with their ids.
      3. Propose. The session learned something; it calls propose_item. Nothing is written yet.
      4. File. The Archivist checks it with your rules, files it, links it, and answers with a receipt.
      5. Remember. Tomorrow, a Codex session asks a related question and finds it.
      
                
      readSearch, open items, list places, find paths.
      propose itemsRead, plus propose new items to the Archivist.
      propose changesAlso propose changes to existing items. Still never a direct write.

      Each agent has its own token and level, and the Brain listens on your computer only. Levels are permissions inside the app, not network security: read the AI notice.

      What it takes, honestly

      • Meaning needs an OpenAI key. Without one, everything still works with words: search, filing by hand, the graph.
      • What you send to a model leaves your computer. The text of a note to make its vector, your question to answer it. Your Brain itself, its files and its history stay in your user folder.
      • Built for one person on one computer. It is a lab, not a server: don't put it on the internet.